File size: 6,219 Bytes
c968fc3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
# This module is from [WeNet](https://github.com/wenet-e2e/wenet).

# ## Citations

# ```bibtex
# @inproceedings{yao2021wenet,
#   title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
#   author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
#   booktitle={Proc. Interspeech},
#   year={2021},
#   address={Brno, Czech Republic },
#   organization={IEEE}
# }

# @article{zhang2022wenet,
#   title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
#   author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
#   journal={arXiv preprint arXiv:2203.15455},
#   year={2022}
# }
#

"""Positonal Encoding Module."""

import math
from typing import Tuple, Union

import torch
import torch.nn.functional as F


class PositionalEncoding(torch.nn.Module):
    """Positional encoding.

    :param int d_model: embedding dim
    :param float dropout_rate: dropout rate
    :param int max_len: maximum input length

    PE(pos, 2i)   = sin(pos/(10000^(2i/dmodel)))
    PE(pos, 2i+1) = cos(pos/(10000^(2i/dmodel)))
    """

    def __init__(
        self,
        d_model: int,
        dropout_rate: float,
        max_len: int = 5000,
        reverse: bool = False,
    ):
        """Construct an PositionalEncoding object."""
        super().__init__()
        self.d_model = d_model
        self.xscale = math.sqrt(self.d_model)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        self.max_len = max_len

        self.pe = torch.zeros(self.max_len, self.d_model)
        position = torch.arange(0, self.max_len, dtype=torch.float32).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, self.d_model, 2, dtype=torch.float32)
            * -(math.log(10000.0) / self.d_model)
        )
        self.pe[:, 0::2] = torch.sin(position * div_term)
        self.pe[:, 1::2] = torch.cos(position * div_term)
        self.pe = self.pe.unsqueeze(0)

    def forward(
        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Add positional encoding.

        Args:
            x (torch.Tensor): Input. Its shape is (batch, time, ...)
            offset (int, torch.tensor): position offset

        Returns:
            torch.Tensor: Encoded tensor. Its shape is (batch, time, ...)
            torch.Tensor: for compatibility to RelPositionalEncoding
        """

        self.pe = self.pe.to(x.device)
        pos_emb = self.position_encoding(offset, x.size(1), False)
        x = x * self.xscale + pos_emb
        return self.dropout(x), self.dropout(pos_emb)

    def position_encoding(
        self, offset: Union[int, torch.Tensor], size: int, apply_dropout: bool = True
    ) -> torch.Tensor:
        """For getting encoding in a streaming fashion

        Attention!!!!!
        we apply dropout only once at the whole utterance level in a none
        streaming way, but will call this function several times with
        increasing input size in a streaming scenario, so the dropout will
        be applied several times.

        Args:
            offset (int or torch.tensor): start offset
            size (int): required size of position encoding

        Returns:
            torch.Tensor: Corresponding encoding
        """
        # How to subscript a Union type:
        #   https://github.com/pytorch/pytorch/issues/69434
        if isinstance(offset, int):
            assert offset + size < self.max_len
            pos_emb = self.pe[:, offset : offset + size]
        elif isinstance(offset, torch.Tensor) and offset.dim() == 0:  # scalar
            assert offset + size < self.max_len
            pos_emb = self.pe[:, offset : offset + size]
        else:  # for batched streaming decoding on GPU
            assert torch.max(offset) + size < self.max_len
            index = offset.unsqueeze(1) + torch.arange(0, size).to(
                offset.device
            )  # B X T
            flag = index > 0
            # remove negative offset
            index = index * flag
            pos_emb = F.embedding(index, self.pe[0])  # B X T X d_model

        if apply_dropout:
            pos_emb = self.dropout(pos_emb)
        return pos_emb


class RelPositionalEncoding(PositionalEncoding):
    """Relative positional encoding module.
    See : Appendix B in https://arxiv.org/abs/1901.02860
    Args:
        d_model (int): Embedding dimension.
        dropout_rate (float): Dropout rate.
        max_len (int): Maximum input length.
    """

    def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000):
        """Initialize class."""
        super().__init__(d_model, dropout_rate, max_len, reverse=True)

    def forward(
        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute positional encoding.
        Args:
            x (torch.Tensor): Input tensor (batch, time, `*`).
        Returns:
            torch.Tensor: Encoded tensor (batch, time, `*`).
            torch.Tensor: Positional embedding tensor (1, time, `*`).
        """
        self.pe = self.pe.to(x.device)
        x = x * self.xscale
        pos_emb = self.position_encoding(offset, x.size(1), False)
        return self.dropout(x), self.dropout(pos_emb)


class NoPositionalEncoding(torch.nn.Module):
    """No position encoding"""

    def __init__(self, d_model: int, dropout_rate: float):
        super().__init__()
        self.d_model = d_model
        self.dropout = torch.nn.Dropout(p=dropout_rate)

    def forward(
        self, x: torch.Tensor, offset: Union[int, torch.Tensor] = 0
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Just return zero vector for interface compatibility"""
        pos_emb = torch.zeros(1, x.size(1), self.d_model).to(x.device)
        return self.dropout(x), pos_emb

    def position_encoding(
        self, offset: Union[int, torch.Tensor], size: int
    ) -> torch.Tensor:
        return torch.zeros(1, size, self.d_model)